These days, hundreds of billions of dollars are invested into companies which claim that they are „close to creating AGI“. Several definitions for AGI are circulating. Let’s take the one from the English Wikipedia:
„Artificial general intelligence (AGI) is a type of artificial intelligence (AI) that matches or surpasses human capabilities across a wide range of cognitive tasks.“
Sounds good! But would it be really useful? I believe rather not and I will explain the reasons for this surprising result in the following.
AGI, according to this definition, is a „drop-in“ replacement for humans. If you own a car factory, you could fire all the workers and replace them by AGIs (embodied in a humanoid robot[1]) and let them do all the work, the human workers did before. By doing this, you could - if you could do it right now - undoubtedly save a lot of money.
But this is not what’s going to happen! Humans are living creatures designed to survive in the wild nature. Even if they are surely the first complex system which was able to manufacture cars, it’s definitely not what they are really good at. And as we will see soon, humans are in fact extremely poor car manufacturers.
The car factory of the future will not contain humanoid robots with hammers in their metal hands. And this is the problem why we overestimate the value of AGI so heavily: AI is not most useful when it does a man’s job! AI is most powerful when it can do things humans are very poor at (or even can’t do at all!).
How then could the car factory of the future look? Instead of the mentioned humanoid robots we should rather imagine a highly complex and intelligent 3D printer. Or a swarm of intelligent insect sized robots. Things beyond your wildest dreams. But it is very clear that there must be ways to create products which are far far more efficient than manufacturing them using human hands and tools.
Now we come to the famous quote attributed to Henry Ford:
„If I had asked people what they wanted, they would have said faster horses.“ - Henry Ford (?)
And we see: it’s exactly the same situation again. Most people simply lack the imagination to realize what an extremely powerful tool AI really is. Good examples for truly significant current advances in AI are AlphaFold (which can predict protein structures from DNA sequences, developed by Google DeepMind) or AI systems enabling complex quantum chemical calculations or predicting the physical properties of new materials. AIs are currently revolutionizing chip design and there are AIs for mathematics which can prove complex theorems. These are all tasks at which humans perform only extremely poorly. Machines can outperform humans by a factor of 1000 or more!
Future AIs will work in a way which will be very hard to understand in detail by humans. They will not communicate with each other using our human language. Instead, they will exchange highly structured data. And, as they will required also such data as input, we will need a human-machine interface which enables us to control them (i.e. generate this structured data from human language prompts).
And this is where large language models (LLMs) come into the picture. While they indeed only mimick human intelligence (using machine intelligence and lots of data produced by humans)[2], they are already now perfectly fit for this task [3].
We therefore really don’t need language models which are „smarter than us“. This would be only a „faster horse“. What companies will really want, is specialized non human-like AIs and robots which outperform humans by several orders of magnitude.
But what about jobs which require a deep understanding of human nature? Things like marketing? Again, the most successful ways to solve marketing problems most probably don’t require a human way of thinking. I’m very sure that in the near future AIs specialized to analyze human psychology will figure out ways to manipulate humans which are beyond anything we can imagine today. Probably some recommender systems of social media platforms are doing this already today. This is, of course, extremely dangerous, but this is not the topic of this blog post. As things are today, if a tool for a task is effective, it will be used.
Like this the economy will more and more transform into a system of specialized networked AI agents. And as fewer and fewer humans will be part of this system, the need for them to communicate with humans will also slowly diminish. Most of the communication will be done by AIs which will exchange complex structured data which is hard to understand for humans.
Therefore: AI is not coming for your job (in the sense that a robot will be doing your job)! It's much more likely that it will make your job simply disappear.
[1] Many experts believe that embodiment will be required to achieve AGI
[2] It is actually extremely difficult (if not impossible) to create true human-like AI (read my 2018 blog post about this question)
[3] It is easy to make the output of LLMs conform to the syntax of any formal language (like JSON or XML)
Image: DALL-E
Follow me on X to get informed about new content on this blog.
